---

name: training-and-configs description: "Guides Nerfstudio method selection, ns-train command construction, typed config overrides, resume behavior, logging, CUDA, and reduced smoke checks." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Training and Configs

Use this route when the task is to train a Nerfstudio model, choose a built-in method, edit a typed config, resume from a run, or build a correct ns-train command.

What this route covers

• Built-in methods such as nerfacto, splatfacto, instant-ngp, vanilla-nerf, tensorf, mipnerf, dnerf, depth-nerfacto, neus, and neus-facto.
• CLI ordering: ns-train {method} [method args] {dataparser} [dataparser args].
• The --data alias, dataparser overrides, viewer/logging options, CUDA/CPU choices, ray batch memory knobs, and multi-GPU flags.
• Resume and checkpoint distinction: --load-dir for model checkpoint loading during training, --load-config for loading a saved config where supported.
• Reduced CPU smoke checks versus real CUDA training.

What this route excludes

• Raw capture conversion and transforms.json validation: use data-preparation.
• Viewer/eval/render/export after a run exists: use visualization-and-export.
• Packaging custom methods/dataparsers: use api-extension.

Read/run these bundled files

• references/model-overview.md for built-in method families and backend notes.
• references/configuration.md for typed config and CLI override rules.
• references/workflows.md for training, resume, logging, and multi-GPU recipes.
• references/troubleshooting.md for CUDA, tyro, OOM, and checkpoint issues.
• scripts/build_ns_train_command.py to construct a command string without launching training.

Safe workflow

1. Validate or identify the dataset directory.
2. Choose a method whose model/data assumptions match the scene and hardware.
3. Use ns-train METHOD --help and, when needed, ns-train METHOD DATAPARSER --help to verify flag ownership.
4. Build the command with method flags before dataparser flags.
5. Run a small CPU or CUDA smoke only when it is explicitly a smoke task; otherwise treat real training as a long GPU job.
6. Preserve the generated config.yml for viewer/eval/render/export handoff.